On 3/12/20 3:53 PM, Zack Weinberg wrote:
On further investigation, there is a concrete reason why autoconf 2.69 prefers SHELL=/bin/bash to SHELL=/bin/dash, namely that dash doesn't support $LINENO, necessitating a complicated and fragile workaround (see _AS_LINENO_PREPARE in lib/m4sugar/m4sh.m4). This has been the behavior for a very long time (since roughly 2008). Moreover, the changed behavior of autoconf trunk was not intentional; it's because commit 2b59b6f8a79b8bf77e178ff4e5aa0ede433d39cf missed a direct use of $as_echo in _AS_DETECT_BETTER_SHELL, causing _AS_DETECT_BETTER_SHELL *not to work at all*! I'm still looking into why tests/testsuite malfunctions when run by dash, but right now I am inclined to say that the bug in _AS_DETECT_BETTER_SHELL should be fixed and then we should independently discuss whether it makes sense to prefer dash to bash. It *would* discourage people from writing bashisms in their configure.ac's and it *is* 30% faster than bash (on my computer, according to "time make check"), but not supporting $LINENO is a big loss.
Here's a thought - now that we know that bash 5 has a severe bug where the presence of a file named '%sn' in the current working directory causes ./configure to fail (for any configure that uses $as-echo), should we teach _AS_DETECT_BETTER_SHELL to catch that bash bug and prefer dash (even with no LINENO support) over bash in that scenario?
-- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org
